From Diffusion to Anomalous Diffusion:A Century After Einstein

M.F. Shlesinger (ONR)A. Blumen (Freiburg)G. Zumofen (ETH)J. Drager (Hamburg)E. Barkai (Notre Dame)R. Metzler (Nordita)S. Denisov (Dresden)O. Flomenbom (TAU)I.M. Sokolov (Berlin)A. Chechkin (Kharkov)A. Lubelski (TAU)

The Diffusion Equation

),(),( 2

2

txPx

KtxPt ∂

∂=

∂∂

A. FickA. Einstein

M. Smoluchowski

Brownian Motion

The Diffusion Equation (1855)Continuity

),(div),( txjtxnt

rrr −=∂∂

+ linear response

=> the diffusion equation

the Green’s function solution

),(),( txPtxn rr →Essentially an equation for the pdf:

(1914,1915,1918)

),(),( txftxn rrrμ+

( )),(),( txntxf rrrμ∇−

),(grad),( txnKtxj rrr−=

),(),( txnKtxnt

rr Δ=∂∂

⎟⎠

⎞⎜⎝

⎛−= −

KtxKttxn d

4exp)4(),(

22/

rr π

“The problem of the random walk”

“Can any of you readers refer me to a work wherein I should find a solution of the following problem, or failing the knowledge of any existing solution provide me with an original one? I should be extremely grateful for the aid in the matter. A man starts from the point O and walks l yards in a straight line; he then turns through any angle whatever and walks another l yards in a second straight line. Herepeats this process n times. Inquire the probability that after n stretches he is at a distance between r and r + δr from his starting point O”.

Karl PearsonNature, 1905

Pearson’s random walk

Emergence of Normal DiffusionEinstein, 1905

Postulates:i) Independent particles

ii) The particle’s motion during two consequent intervals is independent

iii) The displacement during t is s.

For unbiased diffusion:

Moreover, ∞<= ∫∞

∞−

dsss )(22 φλ

)()( ss −= φφ Essentially the Random Walk Model(1880, 1900, 1905×2)

Motion as a sum of small independent increments: ∑=

=N

iistx

1)(

mean free path2/12

is=λ∞<< λ0

mean relaxation time2/12/ vλτ ∝

∞<< τ0τ/tN ≅

the central limit theorem

⎟⎠

⎞⎜⎝

⎛−= −

KtxKttxP

4exp)4(),(

22/1π

τλτ /22 ≡∝ vKwith

ji

N

ii ssNsNstx 2)( 2

2

1

2 +=⎟⎠⎞

⎜⎝⎛= ∑

=

Brownian motion (simple random walk)

; K is the diffusion coefficient

Anomalous diffusion

(a)

Subdiffusion (dispersive)Superdiffusion

(b)

strong anomaly, ultraslow

Aim: creating framework for treating anomaly and strong anomaly in diffusion.

αttr ~)(2 ><

1<α1>α

)(log~)(2 ttr β><

Kttr ~)(2 ><

Physics of Disorder- Subdiffusion

H. Scher and E. Montroll, 1975 in crystalline solids

∫==L

dltlndtd

dtdPtI

0

),()(

In disordered solids (no matter organic or inorganic...)

Dispersion of Contaminants

Diffusion of tracers in fluid flows.

Large scale structures (eddies, jets or convection rolls) dominatethe transport.Example. Experiments in a rapidly rotating annulus (Swinney et al.).

Ordered flow:Levy diffusion(flights and traps)

Weakly turbulent flow:Gaussian diffusion

Superdiffusion

The Flight of the Albatross

Searching for Food

Normal vs. Anomalous Diffusion

Anomalous is Normal...

CTRW Frameworks1. Probability distribution to make a step r in time t.

Single motion event

2. Jump Model

p.d. of being at r at time t.

Fourier-Laplace

3. , decoupling

4. Velocity model:

=),( trψ

)()(),( trptr ψψ =

Probability to move distance r’ atτ τ

5. Possible generalizations:Distribution of velocitiesCoupled modelsDistance dependent velocities

6. ')'()(~),(~ dttvtrtrt∫∞

−Ψ ψδ

Waiting Times

Continuous time random walk (CTRW) model.

CTRW

1. Within the random walk framework:(a) Small changes in step length or time not enough (b) Need for processes on all scales;broad distribution no moments

2.

3. Long tailed distributions temporal (subdiffusion)spatial (superdiffusion)

αKttr ~)(2 ><Simple distributions

Broad distributions

Explanation: The CTRW

)/exp()/exp()(

0

0

TkEEEE

Bii

ii

−=−∝

ττρ

αψ −−∝ 1)( ttThe waiting-time distribution between the two jumpswith 0/ ETkB=α

Diffusion anomalies for 0 < α < 1: the mean waiting time diverges!

Note: The CTRW processes with the power-law waiting times(with 0 < α < 1) are always nonstationary!

The Subordination

PDF of the particle’sposition after n steps

(say, a Gaussian)

Probability to makeexactly n steps up to

the time t

∑∞

=

=0

)(),(),(n

n tnxFtxP χ

)(~)1(11)(~1),(~̂

22 ukuuukf

ψλψ

−−−

=

the limiting form of the characteristic function in Laplace representation

The second moment αψ −−∝ 1)( tt ⇒ αψ Auu −≅ 1)( αttx ∝)(2

Forms of stable distributions (Levy)

10 << α20 << α

αα +∞→ 11~)(

xxPαα +∞→ 1

1~)(x

xP

yxCyxBAyxV coscos)sin(cos),( +++=

Questions:

• Can Anomalous diffusion be described on the same level of description as simple diffusion?

• How does anomalous diffusion modify reactions and first passage times?

• Can we use a generalization of:

),(),( 2

2

txPx

KtxPt ∂

∂=∂∂

Fractional Diffusion Equations• Following scaling arguments one can postulate equationswhich are of non-integer order in time or in space, e.g.

),(),( 2

2

txPx

KtxPt ∂

∂=

∂∂

α

α

or

),(),( 2

2

1

1

txPxt

KtxPt ∂

∂∂∂

=∂∂

−

−

α

α

Such equations allow for :• easier introduction of external forces• introduction of boundary conditions• using the methods of solutions known for “normal” PDEs

Fractional Derivatives

1695 Leibnitz - de l’Hospital

nmnmmn

n

tnm

mtnm

mtdtd −−

−+Γ+Γ

≡−

=)1(

)1()!(

!

Trivial generalization: νμμνμν

ν

νμμ −

−+Γ+Γ

== ttDtdtd

t )1()1(

0

Interesting: νν

ν−

−Γ= tDt )1(

110

This definition is enough to handle the functions which can beexpanded into Taylor series, but obscures the nature of the fractional differentiation operator.

All modern definitions are based on generalizations of the repeated integration formula:

∫ ∫ ∫ ∫−

−− −−

==x

a

y

a

y

a

x

a

nnn

nxa

n

dyyfyxn

dydyyfxfD1 1

)()()!1(

1...)(...)( 11

Its generalization is: The fractional integral

)10()'()'(

)(1)(

0

0 1 <<−Γ

= ∫ −− p

tttf

ptfD

t

tp

ptt

Fractional derivatives may be defined through additional differentiation:

[ ])1()()( )(00

+== −− qntfDdtdtfD qn

ttn

nqtt

Fractional derivatives are nonlocal integral operatorsand are best suited for the description of nonlocalities in space(long jumps) or time (memory effects)

∫ −−Γ∂∂

=− t

t dttfttt

tfD0

10 ')'(

)'(1

)1(1)( α

α

α

)',(),( 2

21

0 txPx

KDtxPt t ∂

∂=

∂∂ −α

The Fractional Fokker-Planck Equation

• Force-free mean squared displacement

• Stationary solution

2sec][ −= ααη

αα

αt

Ktx

)1(2

)( 02

+Γ=><

α

α ηmTkK B= Generalized Einstein-Stokes relation

Not all systems obey the fractional Fokker-Planck Equation

• Fractional Langevin equation

)()()( 02

2

tTtxDmtxdtdm t =+ α

αη

T(t) = Gaussian random noiseα−tTtT ~)0()(

• The corresponding diffusion equation is:

),(),( FBM2

21

FBM txPx

tKtxPt ∂

∂=

∂∂ −α

αα

Local in time

The fractional Brownian motion

Single molecule experiments in proteins

• Fluorescence resonant energy transfer (tens of angstroms).• Photo-induced electron transfer (a few angstroms)

eqXtXtx −= )()(

S. C. Kou and X. S. Xie, PRL (2004)W. Min et al., PRL (2005)

Not all diffusion processes lead to power-law scaling

Prominent examples are:•Simple crossover models in CTRW: Waiting-time distributionswith exponential cutoff•Sinai diffusion: Retarding subdiffusion typically with logarithmicscaling (genuine model corresponds to α = 4).•Retarding superdiffusion: Truncated Lévy flights.

ttx αln)(2 ∝

Although such examples can easily be formulated within theframework of (continuous time) random walks, they can notbe described through the fractional diffusion equations. ⇒ We need generalizations of the FDE scheme.

Distributed-Order Diffusion Eqns.

2

21

0

1 )(xPK

tPpd

∂∂

=∂∂

∫ −β

ββ ββτ

Example: The equation with the distributed-order Caputoderivative in the l.h.s.

The Fourier-Laplace representation of the Green’s function

τττ

KksIsI

sskf 2)(

)(1),(~̂+

= )()()(1

0

βτβτ β psdsI ∫=

shows that a process may be subordinated to a Wiener process:

with

),(4

),(0

4/2

tTedtxPx

ωπω

ωω

∫∞ −

= with )()(),(~ τωτω sIessIsT −=

References:“Beyond Brownian Motion”Physics Today, 49, 33 (1996).

“Strange Kinetics”Nature 363, 31 (1993).

“The Random Walk Guide…”Physics reports 339, 1 (2000).

“Fractional Kinetics”Physics Today 55, 48 (2002).

“The Restaurant at the End of the Random Walk….”Journal Physics A 37, R161 (2004).

“Anomalous Diffusion Spreads its Wings”Physics World 18, 29 August (2005).

Simple Example: A Crossover)()()( 2211 ββδββδβ −+−= BBpTwo different β-values:

with β1 < β22

2

21

1

121

1

0

)()()( βββββ ττβτβτ sBsBpsdsIbb

+== ∫

⎟⎠

⎞⎜⎝

⎛−=

⎭⎬⎫

⎩⎨⎧= −

+−12

212

2

2

11,

2

1-2 2)(

12)( βββββ

βττ

τ tbbEt

bK

ssIKtx sL

From the asymptotics of the generalized Mittag-Leffler functionat small / large negative arguments one gets

The process described is a decelerating subdiffusion

2

2

)1(2

22

2 ββ

τβτ tt

BDx ∝⎟

⎠⎞

⎜⎝⎛

+Γ≈ 1

1

)1(2

11

2 ββ

τβτ tt

BDx ∝⎟

⎠⎞

⎜⎝⎛

+Γ≈

large tsmall t

Example: Sinai-like Diffusion1)( −= ννββpThe Sinai-like models correspond to

( )[ ]⎩⎨⎧

<<+Γ>>

≈= ∫ 1,/1ln/)1(1,ln/

)()()(1

0 ττντττν

βτβτ νβ

sssss

psdsI

The long-time behavior of the pdf (t/τ >> 1):

⎭⎬⎫

⎩⎨⎧

⎟⎠⎞

⎜⎝⎛ +Γ

−⎥⎦

⎤⎢⎣

⎡ +Γ≈

)/(ln||)1(exp

)/(ln)1(

41),( 2/

2/12/1

ττν

τν

τ νν tx

KtDtxP

Note the characteristic tent-like shape of the pdf with

)/(ln)1(

2)(2 τν

τ ν tKtx+Γ

≅

so that

Riemann-Liouville DODE for subdiffusion

∫ ∂∂

=∂

∂ −−1

02

21

01 ),()(),( txP

xDKpd

ttxP

tββτββ

[ ]( )τKksIssIskf

RLRL2)()(

1),(+

=

The Laplace transform of the characteristic function:

with ∫ −=1

0

))(()( βτββ spdsIRL

The solution in a subordination form:

),(4

),(0

4/2

tTedtxP RL

x

ωπω

ωω

∫∞ −

= with )(/

)(1),(~ sI

RLRL e

ssIsT ωω −=

The crossover model for subdiffusion

)()()( 2211 ββδββδβ −+−= BBpTwo different β-values:with β1 < β2

221121

1

0

)()()( βββββ ττβτβτ sBsBpsdsI +== ∫

The process described is an accelerated subdiffusion

1

1

)1(2

11

2 ββ

τβτ tt

BKx ∝⎟

⎠⎞

⎜⎝⎛

+Γ≈

small t

21

)1(2

)1(2)(2)(

2

2

1

112ββ

τβτ

τβτττ ⎟

⎠⎞

⎜⎝⎛

+Γ+⎟

⎠⎞

⎜⎝⎛

+Γ=

⎭⎬⎫

⎩⎨⎧= − tKBtKB

ssILDtx RL

s

2

2

)1(2

2

22 ββ

τβτ ttKBx ∝⎟

⎠⎞

⎜⎝⎛

+Γ≈

large t

Example: A simple crossover

111

1

)(

1)(

++ +≅

−−=Ψ

βα

βα

βαψt

bt

at

tb

tat

Easier to start from cumulativefunctions:

2

11

2

2

)()(

11)(

βα

βα

βα

βαψdtct

tdtct

dtctt

++

≅

+−=Ψ

−−Acceleratingprocess

Slowing downprocess

The fluorescence lifetime of flavin is shortened upon binding to the protein.

Free FLavin in water Flavin bound in FRE

Long and single exponential Shortened and multi-exponential

NADH:Flavin Oxidoreductasefrom E. Coli

Rules of Thumb

The equations with the distributed-order fractional derivative on the “proper” side describe the processes getting more anomalous in the course of time.

The equations with the distributed-order fractional derivative on the “wrong” side describe the processes getting less anomalous in the course of time.

Levy Flights1. Consider a random walker each of whose steps is an identically

distributed random variable with pdf p(x).

2. When can Pn(x), the pdf of being at x after n steps, be the same as p(x)?

3.

4. , Gaussian; (finite)

,

]exp[)( γknkPn −= 20 ≤< γγ−−

∞→

1~)(lim xnxPnx20 << γ

2=γnnx ~)(2 ><

2<γ γkckPnk−

→1~)(lim

0

∞→∂∂−=>=< 2222 /)()()( kkPxxdxPnx=∫ 0knn

Spatial derivatives

Riesz-Weil symmetric derivative:

( )[ ]

⎪⎩

⎪⎨

⎧

=−

≠+−=

∞∞−

1,

1,2/cos2

1

)(αφ

αφφπαφ

αα

α

α

Hdxd

DDx

xdd xx

with H being a Hilbert-transform.

In the Fourier-representation

φφ αα

αˆ||

||ˆ k

xdd

−=⎟⎠

⎞⎜⎝

⎛Φ

Possible positions of derivatives:

Riemann-Liouville derivative on the “wrong” side (r.h.s.)

Riesz-Weyl derivative on the “correct” side (r.h.s.)

Riesz-Weyl derivative on the “wrong” side (l.h.s.)

Caputo derivative on the “correct” side (l.h.s.)

),(),( 2

2

txPx

KtxPt ∂

∂=

∂∂

α

α),(),( 2

2

1

1

txPx

Kt

txPt ∂

∂∂∂

=∂∂

−

−

α

α

),(),( 2

2

txPx

KtxPt β

β

∂∂

=∂∂ ),(),( 2

2

22

22

txPx

KtxPtx ∂

∂=

∂∂

∂∂

− −

−

β

β

Normal forms: Modified forms:

The Fractional Fokker-Planck Equation

• Force-free mean squared displacement

• Stationary solution

2sec][ −= ααη

αα

αt

Ktx

)1(2

)( 02

+Γ=><

α

α ηmTkK B= Generalized Einstein-Stokes relation

Blinking Quantum Dots

Ultraslow Diffusion: The Sinai Model

( ) 2/1max )( xxU Δ∝ΔScaling considerations:

( )( )2/10

max0 exp

)(exp)( xconst

TkxU

xtB

Δ⋅∝⎟⎠

⎞⎜⎝

⎛ Δ≅Δ ττ

Inverting this relation we get: tx 2log∝Δ tx 42 log∝Δ

The PDF of Δx is approximately double-sided exponential

or

ttx 3/42 log)( ∝

Frameworks for anomalous diffusion

1. Generalized diffusion equation

2. Fractional Brownian motion

3. Random velocity fields

4. Self avoiding walks

5. Fractional Fokker Planck Equation (FFPE)

6. Continuous Time Random Walk (CTRW)

7. Sinai diffusion

Coupled modelPossibility to include within the velocity model, interruptions by spatiallocalization (jump model)

)(~ tψ

Waiting Times

Continuous time random walk (CTRW) model.

Subdiffusion (dispersive transport)1, <ααt

1. Jump Model, decoupling

well behaved

no time scale

2. From CTRW:

)(rp10,~)( 1 <<−− αψ αtt

∞>=< t

)()(),( trptr ψψ =

10,~)(2 <<>< ααttr

Subdiffusion (dispersive transport)

3. one dimension

“Stretched” Gaussian

2// αξ tr=

⎩⎨⎧

=−−−

),),exp()exp(

~)(2

21

αδξξξξξ

ξδ -2/(2 large

small b

aaf

1, <ααt

P(r,t)

1. Brownian motion

Gaussian

2. Subdiffusion

“Stretched” Gaussian

2/12 /)exp()( txcf =−= ξξξ ,

ttx ~)(2 ><

2// αξ tr=

⎩⎨⎧

=−−−

),),exp()exp(

~)(2

21

αδξξξξξ

ξδ -2/(2 large

small b

aaf

αttx ~)(2 ><

Equivalence of the forms

• RL-Form • C-form

2

21

0 xWKWDt ∂

∂=−

ββ

2

21

0 xWDK

tW

t ∂∂

=∂

∂ −ββ

)()0,( xxf δ=under the initial condition

Under the Laplace-transform the equations take the same form:

The two forms of fractional spatial equations show their equivalence under the Fourier transform

),(~)0,(),(~2

21 sxp

xKxpssxps

∂∂

=− −β

ββ

Definitions of derivatives:Temporal derivatives

RL derivative

Caputo derivative

Relation between RL- and C-forms: )0()1(

)()( * fttfDtfD tt μ

μμμ

−Γ+=

−

10)'()'(

1')1(

1

0

1*0 <≤

∂∂

−−Γ= ∫− μ

μ μβ

t

t tfttt

dtfD

10)'()'('

)1(1

0

10 <≤

−∂∂

−Γ= ∫− μ

μ μβ

t

t tttfdt

tfD

Under Laplace-transform

{ } )(~)(0 sfstfDL tμμ = { } )0()(~)( 1

*0 fssfstfDL t−−= μμμ

Some Solutions

Free diffusion:Left: Normal diffusionRight: Subdiffusion with α = 1/2

Ornstein-Uhlenbeckprocess: Diffusion in a harmonic potential

I.M. Sokolov, J. Klafter and A. Blumen, Fractional Kinetics , Physics Today, November 2002, p.48

Fluctuating Enzymes

Broad distributions; Long range correlations; Tails

Generalization of the central limit theorem (Levy stable distributions)

The Diffusion Equation (1855)Continuity

),(div),( txjtxnt

rrr −=∂∂

+ linear response

=> the diffusion equation

the Green’s function solution

),(),( txPtxn rr →Essentially an equation for the pdf:

),(grad),( txnKtxj rrr−=

),(),( txnKtxnt

rr Δ=∂∂

⎟⎠

⎞⎜⎝

⎛−= −

KtxKttxn d

4exp)4(),(

22/

rr π

Simple Random Walk

Schematic representation of a Brownian random walk. The walker jumps at each time step to a randomly selected direction.

⎟⎠

⎞⎜⎝

⎛−= −

KtxKttxP

4exp)4(),(

22/1π

τλτ /22 ≡∝ vKwith

2. Anomalous is normal1. Normal is anomalous

Gaussian, Exponential and Pareto Distributions

Levy-Pareto

• Self similarity (fractals)< t > = ∞or< x2(n) > = ∞

• Modifying the 1905 assumptionsτ → ∞λ → ∞

• Memory or “Funicity” (after “Funes the memorious” by Jorge Luis Borges)

Forms of stable distributions (Levy)

10 << α20 << α

αα +∞→ 11~)(

xxPαα +∞→ 1

1~)(x

xP

“I do not believe a word he is saying, but I am afraid that one day I will have to learn it”

Bob Silbey, 1983

Example: A simple crossover

111

1

)(

1)(

++ +≅

−−=Ψ

βα

βα

βαψt

bt

at

tb

tat

Easier to start from cumulativefunctions:

2

11

2

2

)()(

11)(

βα

βα

βα

βαψdtct

tdtct

dtctt

++

≅

+−=Ψ

−−Acceleratingprocess

Slowing downprocess

111 )( ++ +≅ βαβαψt

bt

at2

11

2 )()( βα

βα βαψdtct

tdtct++

≅−−

Slowing down process Accelerating process

βα −− +≅

BtAttn 1)(1

βα DtCttn +≅)(2

a very simple form, a sum of 2 fractional Caputo operators

βα td

tct

~~)(2 +=Φ

a very simple form, a sum of 2 fractional RL-operators

βα tb

tatC +=)(1 βα dtct

tC+

=1)(2

βα tbtat ~~

1)(2 +≈Φ

does not resemble any fractional diff. operator

does not resemble any fractional diff. operator

Superdiffusion and a Wiener Process

The waiting-time distribution cannot be defined.The density of eventsfollows the power-law

),(4

),(0

4/2

tTedtxPx

τπω

ωτ

∫∞ −

=

),,/(),( ββττ ββ −= − tLttT

Subordination form:

with

βρρ −−∝ 1)(p

“I do not believe a word he is saying, but I am afraid that one day I will have to learn it”

Bob Silbey, 1983

”I used a waiting time distribution that had such a long time tail that the mean time of it did not exist. That was the step! Everything fell into place.”

Harvey Scher

Fractional Diffusion Equations (1985, 1989)

• In many physical situations (non-Fickian diffusion)

2/1,)( ≠∝ ααttX

• Following scaling arguments one can postulate (and sometimesderive) fractional equations for anomalous diffusion:

holds

),(),( 2

2

1

1

txPxt

KtxPt ∂

∂∂∂

=∂∂

−

−

α

α

or),(),( 2

2

txPx

KtxPt β

β

∂∂

=∂∂

Such equations allow for • easier introduction of external forces• introduction of boundary conditions• using the methods of solutions known for “normal” PDEs

Reasonable values of the orders of derivatives: 0 < α < 1 for subdiffusion; 0 < β < 1 for superdiffusion.

Congratulations

Bob

Congratulations

Bob

“I do not believe a word he is saying, but I am afraid that one day I will have to learn it”

Bob Silbey, 1983